I don’t think we can extrapolate from current API pricing, but dramatically improving hardware in terms of cost:performance is the underlying reality.

Betting against that you need to assume exponentially more expensive models every year.